ChatGPT bug leaks internal Tasks system prompt and hidden /local automation endpoint
mytrackpadisbroken · reddit · 2026-10-04
A Reddit user found that ChatGPT's Analytics page exposed a hidden chatgpt.com/local/{UUID} endpoint, revealing the full internal system prompt behind ChatGPT Tasks automations, labeled "System instruction (automation runtime; not user content)". The leaked prompt shows the automation runs non-interactively with tools enabled, hides internal metadata from users, must never create new automations mid-run, and should self-disable completed recurring tasks. The UI also showed a "Custom Light" model selector. The user says there's no public documentation of this endpoint anywhere.
More from Models
- o1-pro launched at $150/$600 per million tokens; frontier models seen at $0.40-0.80 by 2027 — AiBreakfast · 2026-10-04
- Brother codes with GPT-5.6 Luna, sparking call to hide old models in Codex — Angaisb_ · 2026-10-04
- OpenAI halves subscription usage limits amid compute crunch, and Anthropic is the beneficiary — kimmonismus · 2026-10-04
- Suspicious output differences suggest Anthropic is quietly testing Haiku 5.5 — legit_api · 2026-10-04
- Dev hammers Opus 5.5 all week and still never hits the usage cap — therealdanvega · 2026-10-04
- Kimi K3 with 1M context now usable in Claude Code via engy inference — richdotca · 2026-10-04